Abstract / Summary
Abstract Background Colorectal polyps are a major source of precancerous lesions in colorectal cancer (CRC). In many population-based screening programs, a major challenge is the efficient triage of high-risk individuals for diagnostic colonoscopy amid limited endoscopic resources. Objective To enrich current screening frameworks, we aimed to develop an accessible, noninvasive risk stratification tool to serve as a digital triage mechanism for colorectal polyps using machine learning (ML) and routinely collected data in China. Methods We conducted a cross-sectional study in Wenzhou, China. A total of 4108 individuals (aged 50‐74 y) who were referred for and accepted colonoscopy following an initial population-based risk assessment (questionnaire and fecal test) between May and November 2021 were included. The dataset was split into training and validation sets, and the synthetic minority oversampling technique (SMOTE) was applied only to the training dataset to address class imbalance. Twenty-one noninvasive predictors (lifestyle, dietary, clinical symptoms, and family history) were selected using the Boruta algorithm and least absolute shrinkage and selection operator (LASSO) regression. Nine ML models were evaluated, with the Shapley Additive Explanations (SHAP) method and local interpretable model–agnostic explanations (LIME) used for model interpretability and feature ranking. Results Among the 9 ML algorithms evaluated, XGBoost (Extreme Gradient Boosting) achieved the highest area under the receiver operating characteristic curve of 0.672, while LightGBM (Light Gradient Boosting Machine) was identified as the optimal model for clinical triage due to its superior recall (0.6503), a key metric for minimizing missed lesions in community screenings. SHAP analysis identified current smoking status, sex, and family history of colorectal polyps as the most influential factors. Notably, the model captured significant nonlinear risk thresholds, such as an age of 50 years and a BMI of 25 kg/m 2 , providing a more granular risk profile than traditional linear models. Conclusions This study provides a scalable, interpretable triage tool to complement existing 2-step CRC screening protocols. By leveraging only noninvasive variables, the LightGBM model enables prioritized referral for colonoscopy, offering a resource-efficient strategy to optimize CRC prevention in resource-limited settings.